Source code for torch_tensorrt.ts._compile_spec
from __future__ import annotations
from copy import deepcopy
from typing import Any, Dict, List, Optional, Set
import torch
import torch_tensorrt._C.ts as _ts_C
from torch_tensorrt import _C, _enums
from torch_tensorrt._Device import Device
from torch_tensorrt._Input import Input
from torch_tensorrt.logging import Level, log
from torch_tensorrt.ts._Input import TorchScriptInput
import tensorrt as trt
def _internal_input_to_torch_class_input(i: _C.Input) -> torch.classes.tensorrt._Input:
clone = torch.classes.tensorrt._Input()
clone._set_min(i.min)
clone._set_opt(i.opt)
clone._set_max(i.max)
clone._set_dtype(i.dtype)
clone._set_tensor_domain(i.tensor_domain)
clone._set_format(i.format)
clone._set_input_is_dynamic(i.input_is_dynamic)
clone._set_explicit_set_dtype(i._explicit_set_dtype)
return clone
def _supported_input_size_type(input_size: Any) -> bool:
if isinstance(input_size, torch.Size):
return True
elif isinstance(input_size, tuple):
return True
elif isinstance(input_size, list):
return True
else:
raise TypeError(
"Input sizes for inputs are required to be a List, tuple or torch.Size or a Dict of three sizes (min, opt, max), found type: "
+ str(type(input_size))
)
def _parse_op_precision(precision: Any) -> _enums.dtype:
if isinstance(precision, torch.dtype):
if precision == torch.int8:
return _enums.dtype.int8
elif precision == torch.half:
return _enums.dtype.half
elif precision == torch.float:
return _enums.dtype.float
else:
raise TypeError(
"Provided an unsupported dtype as operating precision (support: int8, half, float), got: "
+ str(precision)
)
elif isinstance(precision, _enums.dtype):
return precision
else:
raise TypeError(
"Op precision type needs to be specified with a torch.dtype or a torch_tensorrt.dtype, got: "
+ str(type(precision))
)
def _parse_enabled_precisions(precisions: Any) -> Set[_enums.dtype]:
parsed_precisions = set()
if any(isinstance(precisions, type) for type in [list, tuple, set]):
for p in precisions:
parsed_precisions.add(_parse_op_precision(p))
else:
parsed_precisions.add(_parse_op_precision(precisions))
return parsed_precisions
def _parse_device_type(device: Any) -> _enums.DeviceType:
if isinstance(device, torch.device):
if device.type == "cuda":
return _C.DeviceType.gpu
else:
ValueError(
"Got a device type other than GPU or DLA (type: "
+ str(device.type)
+ ")"
)
elif isinstance(device, _C.DeviceType):
return device
elif isinstance(device, trt.DeviceType):
if device == trt.DeviceType.DLA:
return _C.DeviceType.DLA
return _C.DeviceType.GPU
elif isinstance(device, str):
if device == "gpu" or device == "GPU":
return _C.DeviceType.GPU
elif device == "dla" or device == "DLA":
return _C.DeviceType.DLA
else:
ValueError(
"Got a device type other than GPU or DLA (type: " + str(device) + ")"
)
else:
raise TypeError(
"Device specification must be of type torch.device, string or torch_tensorrt.DeviceType, but got: "
+ str(type(device))
)
def _parse_device(device_info: Any) -> _C.Device:
if isinstance(device_info, dict):
info = _C.Device()
if "device_type" not in device_info:
raise KeyError("Device type is required parameter")
else:
info.device_type = _parse_device_type(device_info["device_type"])
if "gpu_id" in device_info:
assert isinstance(device_info["gpu_id"], int)
info.gpu_id = device_info["gpu_id"]
if "dla_core" in device_info:
assert isinstance(device_info["dla_core"], int)
info.dla_core = device_info["dla_core"]
if "allow_gpu_fallback" in device_info:
assert isinstance(device_info["allow_gpu_fallback"], bool)
info.allow_gpu_fallback = device_info["allow_gpu_fallback"]
return info
elif isinstance(device_info, Device):
return device_info._to_internal()
elif isinstance(device_info, torch.device):
return (Device._from_torch_device(device_info))._to_internal()
else:
raise ValueError(
"Unsupported data for device specification. Expected either a dict, torch_tensorrt.Device or torch.Device"
)
def _parse_torch_fallback(fallback_info: Dict[str, Any]) -> _ts_C.TorchFallback:
info = _ts_C.TorchFallback()
if "enabled" not in fallback_info:
raise KeyError("Enabled is required parameter")
else:
assert isinstance(fallback_info["enabled"], bool)
info.enabled = fallback_info["enabled"]
if "min_block_size" in fallback_info:
assert isinstance(fallback_info["min_block_size"], int)
info.min_block_size = fallback_info["min_block_size"]
if "forced_fallback_ops" in fallback_info:
assert isinstance(fallback_info["forced_fallback_ops"], list)
info.forced_fallback_operators = fallback_info["forced_fallback_ops"]
if "forced_fallback_modules" in fallback_info:
assert isinstance(fallback_info["forced_fallback_modules"], list)
info.forced_fallback_modules = fallback_info["forced_fallback_modules"]
return info
def _parse_input_signature(input_signature: Any, depth: int = 0) -> Any:
if depth > 2:
raise AssertionError(
"Input nesting depth exceeds max supported depth, use 1 level: [A, B], or 2 level: [A, (B, C)]"
)
if isinstance(input_signature, tuple):
input_list = []
for item in input_signature:
input = _parse_input_signature(item, depth + 1)
input_list.append(input)
return tuple(input_list)
elif isinstance(input_signature, list):
input_list = []
for item in input_signature:
input = _parse_input_signature(item, depth + 1)
input_list.append(input)
return input_list
elif isinstance(input_signature, (Input, torch.Tensor)):
i = (
Input.from_tensor(input_signature)
if isinstance(input_signature, torch.Tensor)
else input_signature
)
if not i.is_trt_dtype():
raise TypeError(
"Using non-TRT input types with input_signature is not currently "
+ "supported. Please specify inputs individually to use "
+ "non-TRT types."
)
ts_i = i
if i.shape_mode == Input._ShapeMode.STATIC:
ts_i = TorchScriptInput(shape=i.shape, dtype=i.dtype, format=i.format)
elif i.shape_mode == Input._ShapeMode.DYNAMIC:
if isinstance(i.shape, dict):
ts_i = TorchScriptInput(
min_shape=i.shape["min_shape"],
opt_shape=i.shape["opt_shape"],
max_shape=i.shape["max_shape"],
dtype=i.dtype,
format=i.format,
)
else:
raise ValueError(
f"Input set as dynamic, expected dictionary of shapes but found {i.shape}"
)
else:
raise ValueError(
"Invalid shape mode detected for input while parsing the input_signature"
)
clone = _internal_input_to_torch_class_input(ts_i._to_internal())
return clone
else:
raise KeyError(
"Input signature contains an unsupported type {}".format(
type(input_signature)
)
)
def _parse_compile_spec(compile_spec_: Dict[str, Any]) -> _ts_C.CompileSpec:
# TODO: Use deepcopy to support partial compilation of collections
compile_spec = deepcopy(compile_spec_)
info = _ts_C.CompileSpec()
if len(compile_spec["inputs"]) > 0:
if not all(
isinstance(i, (torch.Tensor, Input)) for i in compile_spec["inputs"]
):
raise KeyError(
"Input specs should be either torch_tensorrt.Input or torch.Tensor, found types: {}".format(
[type(i) for i in compile_spec["inputs"]]
)
)
inputs = [
Input.from_tensor(i) if isinstance(i, torch.Tensor) else i
for i in compile_spec["inputs"]
]
ts_inputs = []
for i in inputs:
if i.shape_mode == Input._ShapeMode.STATIC:
ts_inputs.append(
TorchScriptInput(
shape=i.shape, dtype=i.dtype, format=i.format
)._to_internal()
)
elif i.shape_mode == Input._ShapeMode.DYNAMIC:
ts_inputs.append(
TorchScriptInput(
min_shape=i.shape["min_shape"],
opt_shape=i.shape["opt_shape"],
max_shape=i.shape["max_shape"],
dtype=i.dtype,
format=i.format,
)._to_internal()
)
info.inputs = ts_inputs
elif compile_spec["input_signature"] is not None:
log(
Level.Warning,
"Input signature parsing is an experimental feature, behavior and APIs may change",
)
signature = _parse_input_signature(compile_spec["input_signature"])
info.input_signature = _C.InputSignature(signature) # py_object
else:
raise KeyError(
'Module input definitions are requried to compile module. Provide a list of torch_tensorrt.Input keyed to "inputs" in the compile spec'
)
if "enabled_precisions" in compile_spec:
info.enabled_precisions = _parse_enabled_precisions(
compile_spec["enabled_precisions"]
)
if "calibrator" in compile_spec and compile_spec["calibrator"]:
info.ptq_calibrator = compile_spec["calibrator"]
if "sparse_weights" in compile_spec:
assert isinstance(compile_spec["sparse_weights"], bool)
info.sparse_weights = compile_spec["sparse_weights"]
if "disable_tf32" in compile_spec:
assert isinstance(compile_spec["disable_tf32"], bool)
info.disable_tf32 = compile_spec["disable_tf32"]
if "refit" in compile_spec:
assert isinstance(compile_spec["refit"], bool)
info.refit = compile_spec["refit"]
if "debug" in compile_spec:
assert isinstance(compile_spec["debug"], bool)
info.debug = compile_spec["debug"]
if "allow_shape_tensors" in compile_spec:
assert isinstance(compile_spec["allow_shape_tensors"], bool)
info.allow_shape_tensors = compile_spec["allow_shape_tensors"]
if "device" in compile_spec:
info.device = _parse_device(compile_spec["device"])
if "capability" in compile_spec:
assert isinstance(compile_spec["capability"], _enums.EngineCapability)
info.capability = compile_spec["capability"]
if "num_avg_timing_iters" in compile_spec:
assert type(compile_spec["num_avg_timing_iters"]) is int
info.num_avg_timing_iters = compile_spec["num_avg_timing_iters"]
if "workspace_size" in compile_spec:
assert type(compile_spec["workspace_size"]) is int
info.workspace_size = compile_spec["workspace_size"]
if "dla_sram_size" in compile_spec:
assert type(compile_spec["dla_sram_size"]) is int
info.dla_sram_size = compile_spec["dla_sram_size"]
if "dla_local_dram_size" in compile_spec:
assert type(compile_spec["dla_local_dram_size"]) is int
info.dla_local_dram_size = compile_spec["dla_local_dram_size"]
if "dla_global_dram_size" in compile_spec:
assert type(compile_spec["dla_global_dram_size"]) is int
info.dla_global_dram_size = compile_spec["dla_global_dram_size"]
if "truncate_long_and_double" in compile_spec:
assert type(compile_spec["truncate_long_and_double"]) is bool
info.truncate_long_and_double = compile_spec["truncate_long_and_double"]
if "torch_fallback" in compile_spec:
info.torch_fallback = _parse_torch_fallback(compile_spec["torch_fallback"])
log(Level.Debug, str(info))
return info
[docs]def TensorRTCompileSpec(
inputs: Optional[List[torch.Tensor | Input]] = None,
input_signature: Optional[Any] = None,
device: torch.device | Device = Device._current_device(),
disable_tf32: bool = False,
sparse_weights: bool = False,
enabled_precisions: Optional[Set[torch.dtype | _enums.dtype]] = None,
refit: bool = False,
debug: bool = False,
capability: _enums.EngineCapability = _enums.EngineCapability.default,
num_avg_timing_iters: int = 1,
workspace_size: int = 0,
dla_sram_size: int = 1048576,
dla_local_dram_size: int = 1073741824,
dla_global_dram_size: int = 536870912,
truncate_long_and_double: bool = False,
calibrator: object = None,
allow_shape_tensors: bool = False,
) -> torch.classes.tensorrt.CompileSpec:
"""Utility to create a formated spec dictionary for using the PyTorch TensorRT backend
Keyword Args:
inputs (List[Union(torch_tensorrt.Input, torch.Tensor)]): **Required** List of specifications of input shape, dtype and memory layout for inputs to the module. This argument is required. Input Sizes can be specified as torch sizes, tuples or lists. dtypes can be specified using
torch datatypes or torch_tensorrt datatypes and you can use either torch devices or the torch_tensorrt device type enum
to select device type. ::
input=[
torch_tensorrt.Input((1, 3, 224, 224)), # Static NCHW input shape for input #1
torch_tensorrt.Input(
min_shape=(1, 224, 224, 3),
opt_shape=(1, 512, 512, 3),
max_shape=(1, 1024, 1024, 3),
dtype=torch.int32
format=torch.channel_last
), # Dynamic input shape for input #2
torch.randn((1, 3, 224, 244)) # Use an example tensor and let torch_tensorrt infer settings
]
device (Union(torch_tensorrt.Device, torch.device, dict)): Target device for TensorRT engines to run on ::
device=torch_tensorrt.Device("dla:1", allow_gpu_fallback=True)
disable_tf32 (bool): Force FP32 layers to use traditional as FP32 format vs the default behavior of rounding the inputs to 10-bit mantissas before multiplying, but accumulates the sum using 23-bit mantissas
sparse_weights (bool): Enable sparsity for convolution and fully connected layers.
enabled_precision (Set(Union(torch.dtype, torch_tensorrt.dtype))): The set of datatypes that TensorRT can use when selecting kernels
refit (bool): Enable refitting
debug (bool): Enable debuggable engine
capability (torch_tensorrt.EngineCapability): Restrict kernel selection to safe gpu kernels or safe dla kernels
num_avg_timing_iters (int): Number of averaging timing iterations used to select kernels
workspace_size (int): Maximum size of workspace given to TensorRT
truncate_long_and_double (bool): Truncate weights provided in int64 or double (float64) to int32 and float32
calibrator (Union(torch_tensorrt._C.IInt8Calibrator, tensorrt.IInt8Calibrator)): Calibrator object which will provide data to the PTQ system for INT8 Calibration
allow_shape_tensors: (Experimental) Allow aten::size to output shape tensors using IShapeLayer in TensorRT
Returns:
torch.classes.tensorrt.CompileSpec: List of methods and formated spec objects to be provided to ``torch._C._jit_to_tensorrt``
"""
compile_spec = {
"inputs": inputs if inputs is not None else [],
# "input_signature": input_signature,
"device": device,
"disable_tf32": disable_tf32, # Force FP32 layers to use traditional as FP32 format vs the default behavior of rounding the inputs to 10-bit mantissas before multiplying, but accumulates the sum using 23-bit mantissas
"sparse_weights": sparse_weights, # Enable sparsity for convolution and fully connected layers.
"enabled_precisions": enabled_precisions
if enabled_precisions is not None
else set(), # Enabling FP16 kernels
"refit": refit, # enable refit
"debug": debug, # enable debuggable engine
"capability": capability, # Restrict kernel selection to safe gpu kernels or safe dla kernels
"num_avg_timing_iters": num_avg_timing_iters, # Number of averaging timing iterations used to select kernels
"workspace_size": workspace_size, # Maximum size of workspace given to TensorRT
"dla_sram_size": dla_sram_size, # Fast software managed RAM used by DLA to communicate within a layer.
"dla_local_dram_size": dla_local_dram_size, # Host RAM used by DLA to share intermediate tensor data across operations
"dla_global_dram_size": dla_global_dram_size, # Host RAM used by DLA to store weights and metadata for execution
"calibrator": calibrator,
"truncate_long_and_double": truncate_long_and_double,
"allow_shape_tensors": allow_shape_tensors,
}
parsed_spec = _parse_compile_spec(compile_spec)
backend_spec = torch.classes.tensorrt.CompileSpec()
if input_signature is not None:
raise ValueError(
"Input signature parsing is not currently supported in the TorchScript backend integration"
)
for i in parsed_spec.inputs:
clone = _internal_input_to_torch_class_input(i)
backend_spec._append_input(clone)
d = torch.classes.tensorrt._Device()
d._set_device_type(int(parsed_spec.device.device_type))
d._set_gpu_id(parsed_spec.device.gpu_id)
d._set_dla_core(parsed_spec.device.dla_core)
d._set_allow_gpu_fallback(parsed_spec.device.allow_gpu_fallback)
if parsed_spec.torch_fallback.enabled:
raise RuntimeError(
"Partial module compilation is not currently supported via the PyTorch TensorRT backend. If you need partial compilation, use torch_tensorrt.compile"
)
torch_fallback = torch.classes.tensorrt._TorchFallback()
torch_fallback._set_enabled(parsed_spec.torch_fallback.enabled)
torch_fallback._set_min_block_size(parsed_spec.torch_fallback.min_block_size)
torch_fallback._set_forced_fallback_operators(
parsed_spec.torch_fallback.forced_fallback_operators
)
torch_fallback._set_forced_fallback_modules(
parsed_spec.torch_fallback.forced_fallback_modules
)
backend_spec._set_device(d)
backend_spec._set_torch_fallback(torch_fallback)
backend_spec._set_precisions([int(i) for i in parsed_spec.enabled_precisions])
backend_spec._set_disable_tf32(parsed_spec.disable_tf32)
backend_spec._set_refit(parsed_spec.refit)
backend_spec._set_debug(parsed_spec.debug)
backend_spec._set_refit(parsed_spec.refit)
backend_spec._set_capability(int(parsed_spec.capability))
backend_spec._set_num_avg_timing_iters(parsed_spec.num_avg_timing_iters)
backend_spec._set_workspace_size(parsed_spec.workspace_size)
backend_spec._set_dla_sram_size(parsed_spec.dla_sram_size)
backend_spec._set_dla_local_dram_size(parsed_spec.dla_local_dram_size)
backend_spec._set_dla_global_dram_size(parsed_spec.dla_global_dram_size)
backend_spec._set_truncate_long_and_double(parsed_spec.truncate_long_and_double)
backend_spec._set_allow_shape_tensors(parsed_spec.allow_shape_tensors)
backend_spec._set_ptq_calibrator(parsed_spec._get_calibrator_handle())
return backend_spec
